souravrs999 / License-Plate-Recognition

A pytorch implementation of a darkent trained yolov4-tiny model that can detect number plates and helmets if a number plate is detected it is passed through an OCR to recognize the number
MIT License
73 stars 17 forks source link

Which datasets that the LP detector was trained on? #1

Open dzngo opened 3 years ago

dzngo commented 3 years ago

Firstly, thank you for your amazing project, your pretrained tiny-yolov4 model has helped me much on my personal project. I upvoted this repos. However, I would like to ask you which image datasets that you used for the training of the tiny-yolov4 for detecting the licence plate.

Looking forward for your answer.

souravrs999 commented 3 years ago

Thank you and glad this repo was useful for someone, the model was actually trained on a mix of googles open image data-set and custom annoted datasets of normal images consisting of number plates for 2 wheeler's as well, because google's open image data-set only contained the images of cars and trucks but lacked any for smaller vehicles. This project was also made to be deployable on a Jetson nano attached to the dash cam of a car for Indian road condition's hence the custom datasets to include proper detection's for autorickshaw's, bikes and scooters as well. Hope this explanation was clear.

dzngo commented 3 years ago

Thank you for your reply.

It would be grateful if you could share the datasets. In terms of Nano Jetson, did you convert the yolov4-tiny to tensorRT engine plan? If yes, could you explain how you did that and what is the total FPS when the model is deployed on Nano Jetson?

Looking forward for your answer.

souravrs999 commented 3 years ago

Hi again, the complete dataset is uploaded to google driver i can share the link to you if it's ok, I did not convert the model to TensorRT yet i haven't got the time do it yet as i'm not working on it at the moment but i might in the future, The fps was under 10-12 ish when deployed on a nano i know it's not great but it's better than nothing.

Here's the link to the dataset as well as all the accompanying resources hope this helps. https://drive.google.com/drive/folders/1dBYs9abE26ceC_boi75z7qh67KnKb-Pp?usp=sharing

dzngo commented 3 years ago

Thank you for your thoroughness. Btw, actually, 10-12 FPS on Nano Jetson without converting into tensorRT is really good.

Looking forward to your future works.

souravrs999 commented 3 years ago

Thank you for your kind words and good luck on all your future endeavors.

ozett commented 2 years ago

i did not find the images for the dataset quickly, so may i ask this as a question: is the recognition trained on some international license plates? does it recognise license-plates for euopean countries?

thx

souravrs999 commented 2 years ago

i did not find the images for the dataset quickly, so may i ask this as a question: is the recognition trained on some international license plates? does it recognise license-plates for euopean countries?

thx

The model was trained on Google's open image dataset and it did contain license plates of multiple countries including eu